🤖 AI benchmark: hit-rate of 7 models
Seven external AI models (Hermes contour) independently analyze the same matches — predicting the outcome (1X2), total (Over/Under), both teams to score (BTTS) and the exact score. Here we honestly compare their predictions against the real result after the final whistle and combine everything into a single accuracy rating. An informational and analytical snapshot, not betting advice.
Leaderboard · Hockey
| Model | N (settled) | 1X2 | Double chance (1X) | Total goals | BTTS | Exact score | Composite accuracy |
|---|---|---|---|---|---|---|---|
 Google AI
|
11 | 81.8%(9/11) | 81.8%(9/11) | 54.5%(6/11) | 72.7%(8/11) | 0.0%(0/11) | 52.3%(23/44) |
 Claude
|
10 | 60.0%(6/10) | 60.0%(6/10) | 60.0%(6/10) | 70.0%(7/10) | 0.0%(0/10) | 47.5%(19/40) |
 DeepSeek
|
0 | — | — | — | — | — | — |
 ChatGPT
|
0 | — | — | — | — | — | — |
 Qwen
|
0 | — | — | — | — | — | — |
 Kimi
|
0 | — | — | — | — | — | — |
 GLM 5.2
|
0 | — | — | — | — | — | — |
grey — sample <5, not representative; «—» — the model has not made a settled prediction yet.
Double chance (1X) — the same pick counts as a win if the chosen side won or the match drew. Of the 1X2 losses in football/hockey: 0 draws, 6 underdog (total settled 1X2 picks in these sports: 21, double chance combined 71.4% (15/21)). The models almost always take the favorite and don't bet on a draw — double chance shows how many bets are eaten specifically by draws.
Composite accuracy — the share of correct predictions across all shown markets together: (sum of correct picks) ÷ (sum of all settled picks) across the markets 1X2 + Total goals + BTTS + Exact score. Each market-pick weighs equally. This is hit-rate, not profitability — for money/ROI by model see /ai-agent. Total: a push (score exactly on the line) is excluded from the denominator. BTTS is checked against whether both teams scored. «Exact score» — the full final score was guessed correctly (H and A matched); predictions with no recognized score do not count toward the denominator. Double chance (1X): a pick counts as a win if the chosen side won OR it was a draw — it accounts for frequent draws that «eat» bets on the favorite. This metric is informational and is not included in composite accuracy.
Composite model rating · all markets · Hockey
Bar height = the model's composite accuracy across all applicable markets on the current sample. Sorted from best to worst.
Bars are AI models by version; grey/dimmed — sample <5, not representative. The snapshot is informational, not betting advice.
Accuracy by market · Hockey
Where each model is strong: one mini-bar per applicable market, with the percentage and (hits/sample).
Google AI
Composite 52.3%
Claude
Composite 47.5%
DeepSeek
Composite —
ChatGPT
Composite —
Qwen
Composite —
Kimi
Composite —
GLM 5.2
Composite —
The model's favorite by 1X2 = the max of P1/X/P2 in its probabilities; for sports without a draw (tennis, volleyball, etc.) the «X» option doesn't participate. grey — sample <5, not representative. Not betting advice.